Update README.md
Browse files
README.md
CHANGED
|
@@ -21,7 +21,7 @@
|
|
| 21 |
|
| 22 |
### What is ÈwéBench?
|
| 23 |
|
| 24 |
-
ÈwéBench is the **first standardized benchmark** for evaluating Large Language Models (LLMs) on the **Ewe language** (ɛʋɛgbɛ)
|
| 25 |
|
| 26 |
Unlike generic multilingual benchmarks that treat African languages as afterthoughts, ÈwéBench is **designed from the ground up** for Ewe, with culturally relevant tests, native speaker validation, and evaluation criteria that understand Ewe's unique linguistic features (tonality, agglutination, proverbs).
|
| 27 |
|
|
@@ -38,11 +38,11 @@ Unlike generic multilingual benchmarks that treat African languages as afterthou
|
|
| 38 |
|---------|-------------|
|
| 39 |
| **10 categories** | From linguistic comprehension to agentic capabilities |
|
| 40 |
| **107 tests** | Manually crafted, culturally grounded |
|
| 41 |
-
| **Weighted scoring** | ÈwéScore
|
| 42 |
| **Any model** | Works with any OpenAI-compatible API (local or cloud) |
|
| 43 |
| **CLI & API** | Run from terminal or integrate into CI/CD |
|
| 44 |
| **Leaderboard** | Track and compare model progress |
|
| 45 |
-
| **Presets** | One-command evaluation for
|
| 46 |
|
| 47 |
### Quick Start
|
| 48 |
|
|
@@ -55,7 +55,7 @@ cd EweBench
|
|
| 55 |
pip install -r requirements.txt
|
| 56 |
|
| 57 |
# Run with a preset
|
| 58 |
-
python run_benchmark.py --preset
|
| 59 |
|
| 60 |
# Run with a custom endpoint
|
| 61 |
python run_benchmark.py --endpoint http://localhost:11434/v1/chat/completions \
|
|
@@ -84,7 +84,7 @@ python run_benchmark.py --leaderboard
|
|
| 84 |
| 10 | Robustness | 10 | 5% | Consistency under adversarial inputs |
|
| 85 |
| | **Total** | **107** | **100%** | |
|
| 86 |
|
| 87 |
-
### Scoring
|
| 88 |
|
| 89 |
The **ÈwéScore** is a single number (0-100) representing overall Ewe language capability:
|
| 90 |
|
|
@@ -93,12 +93,12 @@ The **ÈwéScore** is a single number (0-100) representing overall Ewe language
|
|
| 93 |
```
|
| 94 |
|
| 95 |
Each test is scored 0.0-1.0 using evaluation methods:
|
| 96 |
-
- **exact_match**
|
| 97 |
-
- **keywords**
|
| 98 |
-
- **multiple_choice**
|
| 99 |
-
- **format**
|
| 100 |
-
- **ewe_quality**
|
| 101 |
-
- **composite**
|
| 102 |
|
| 103 |
**Passing threshold**: A test is "passed" if score ≥ 0.7
|
| 104 |
|
|
@@ -128,7 +128,7 @@ POST /v1/chat/completions
|
|
| 128 |
```
|
| 129 |
|
| 130 |
**Tested providers:**
|
| 131 |
-
-
|
| 132 |
- Google Gemini (OpenAI-compatible endpoint)
|
| 133 |
- Ollama (local)
|
| 134 |
- vLLM (local)
|
|
@@ -140,7 +140,7 @@ POST /v1/chat/completions
|
|
| 140 |
|
| 141 |
### Qu'est-ce qu'ÈwéBench ?
|
| 142 |
|
| 143 |
-
ÈwéBench est le **premier benchmark standardisé** pour évaluer les grands modèles de langage (LLMs) sur la **langue Ewe** (ɛʋɛgbɛ)
|
| 144 |
|
| 145 |
Contrairement aux benchmarks multilingues génériques qui traitent les langues africaines comme des détails, ÈwéBench est **conçu de zéro** pour l'Ewe, avec des tests culturellement pertinents, une validation par des locuteurs natifs, et des critères d'évaluation qui comprennent les particularités linguistiques de l'Ewe (tonalité, agglutination, proverbes).
|
| 146 |
|
|
@@ -161,22 +161,22 @@ cd EweBench
|
|
| 161 |
# Installer les dépendances
|
| 162 |
pip install -r requirements.txt
|
| 163 |
|
| 164 |
-
# Configurer (optionnel
|
| 165 |
cp .env.example .env
|
| 166 |
# Ajouter vos clés API dans .env
|
| 167 |
|
| 168 |
# Lancer avec un preset
|
| 169 |
-
python run_benchmark.py --preset
|
| 170 |
|
| 171 |
# Lancer sur un modèle local
|
| 172 |
python run_benchmark.py --endpoint http://localhost:11434/v1/chat/completions \
|
| 173 |
--model yawo-v10 --verbose
|
| 174 |
|
| 175 |
# Évaluer une seule catégorie
|
| 176 |
-
python run_benchmark.py --preset
|
| 177 |
|
| 178 |
# Comparer deux modèles
|
| 179 |
-
python run_benchmark.py --compare results/
|
| 180 |
|
| 181 |
# Voir le classement
|
| 182 |
python run_benchmark.py --leaderboard
|
|
@@ -198,7 +198,7 @@ python run_benchmark.py --leaderboard
|
|
| 198 |
| 10 | Robustesse | 10 | 5% | Cohérence face aux entrées adverses |
|
| 199 |
| | **Total** | **107** | **100%** | |
|
| 200 |
|
| 201 |
-
### Scoring
|
| 202 |
|
| 203 |
L'**ÈwéScore** est un nombre unique (0-100) représentant la capacité globale en Ewe :
|
| 204 |
|
|
@@ -214,7 +214,7 @@ L'**ÈwéScore** est un nombre unique (0-100) représentant la capacité globale
|
|
| 214 |
|
| 215 |
| # | Model | ÈwéScore | Tests Passed | Date |
|
| 216 |
|---|-------|----------|--------------|------|
|
| 217 |
-
| 🥇 | *En attente de soumissions* |
|
| 218 |
|
| 219 |
> **Soumettre vos résultats** : Exécutez le benchmark, puis ouvrez une PR avec votre fichier de résultats dans `results/`.
|
| 220 |
|
|
@@ -258,16 +258,16 @@ EweBench/
|
|
| 258 |
We welcome contributions! See [docs/CONTRIBUTING.md](docs/CONTRIBUTING.md) for details.
|
| 259 |
|
| 260 |
Ways to contribute:
|
| 261 |
-
- **Add tests**
|
| 262 |
-
- **Validate translations**
|
| 263 |
-
- **Submit results**
|
| 264 |
-
- **Report issues**
|
| 265 |
|
| 266 |
---
|
| 267 |
|
| 268 |
## License
|
| 269 |
|
| 270 |
-
**CC BY-NC 4.0**
|
| 271 |
|
| 272 |
- ✅ Free to use for research, education, and evaluation
|
| 273 |
- ✅ Free to modify and redistribute (with attribution)
|
|
@@ -291,8 +291,8 @@ Ways to contribute:
|
|
| 291 |
|
| 292 |
<div align="center">
|
| 293 |
|
| 294 |
-
**Created by [Joel Elisée ADZONYA](https://joel.adzonya.strivenew.com)
|
| 295 |
|
| 296 |
-
*L'IA au service des langues africaines*
|
| 297 |
|
| 298 |
</div>
|
|
|
|
| 21 |
|
| 22 |
### What is ÈwéBench?
|
| 23 |
|
| 24 |
+
ÈwéBench is the **first standardized benchmark** for evaluating Large Language Models (LLMs) on the **Ewe language** (ɛʋɛgbɛ) a Kwa language spoken by ~7 million people in Togo and Ghana.
|
| 25 |
|
| 26 |
Unlike generic multilingual benchmarks that treat African languages as afterthoughts, ÈwéBench is **designed from the ground up** for Ewe, with culturally relevant tests, native speaker validation, and evaluation criteria that understand Ewe's unique linguistic features (tonality, agglutination, proverbs).
|
| 27 |
|
|
|
|
| 38 |
|---------|-------------|
|
| 39 |
| **10 categories** | From linguistic comprehension to agentic capabilities |
|
| 40 |
| **107 tests** | Manually crafted, culturally grounded |
|
| 41 |
+
| **Weighted scoring** | ÈwéScore single metric, weighted by category importance |
|
| 42 |
| **Any model** | Works with any OpenAI-compatible API (local or cloud) |
|
| 43 |
| **CLI & API** | Run from terminal or integrate into CI/CD |
|
| 44 |
| **Leaderboard** | Track and compare model progress |
|
| 45 |
+
| **Presets** | One-command evaluation for Gemini, local models |
|
| 46 |
|
| 47 |
### Quick Start
|
| 48 |
|
|
|
|
| 55 |
pip install -r requirements.txt
|
| 56 |
|
| 57 |
# Run with a preset
|
| 58 |
+
python run_benchmark.py --preset model --verbose
|
| 59 |
|
| 60 |
# Run with a custom endpoint
|
| 61 |
python run_benchmark.py --endpoint http://localhost:11434/v1/chat/completions \
|
|
|
|
| 84 |
| 10 | Robustness | 10 | 5% | Consistency under adversarial inputs |
|
| 85 |
| | **Total** | **107** | **100%** | |
|
| 86 |
|
| 87 |
+
### Scoring ÈwéScore
|
| 88 |
|
| 89 |
The **ÈwéScore** is a single number (0-100) representing overall Ewe language capability:
|
| 90 |
|
|
|
|
| 93 |
```
|
| 94 |
|
| 95 |
Each test is scored 0.0-1.0 using evaluation methods:
|
| 96 |
+
- **exact_match** Normalized string comparison
|
| 97 |
+
- **keywords** Presence of expected Ewe keywords
|
| 98 |
+
- **multiple_choice** QCM answer detection
|
| 99 |
+
- **format** Output format compliance (markdown, function_call, etc.)
|
| 100 |
+
- **ewe_quality** Heuristic Ewe linguistic quality (character usage, vocabulary, structure)
|
| 101 |
+
- **composite** Weighted combination of multiple methods
|
| 102 |
|
| 103 |
**Passing threshold**: A test is "passed" if score ≥ 0.7
|
| 104 |
|
|
|
|
| 128 |
```
|
| 129 |
|
| 130 |
**Tested providers:**
|
| 131 |
+
- Openai SDK
|
| 132 |
- Google Gemini (OpenAI-compatible endpoint)
|
| 133 |
- Ollama (local)
|
| 134 |
- vLLM (local)
|
|
|
|
| 140 |
|
| 141 |
### Qu'est-ce qu'ÈwéBench ?
|
| 142 |
|
| 143 |
+
ÈwéBench est le **premier benchmark standardisé** pour évaluer les grands modèles de langage (LLMs) sur la **langue Ewe** (ɛʋɛgbɛ) une langue Kwa parlée par ~7 millions de personnes au Togo et au Ghana.
|
| 144 |
|
| 145 |
Contrairement aux benchmarks multilingues génériques qui traitent les langues africaines comme des détails, ÈwéBench est **conçu de zéro** pour l'Ewe, avec des tests culturellement pertinents, une validation par des locuteurs natifs, et des critères d'évaluation qui comprennent les particularités linguistiques de l'Ewe (tonalité, agglutination, proverbes).
|
| 146 |
|
|
|
|
| 161 |
# Installer les dépendances
|
| 162 |
pip install -r requirements.txt
|
| 163 |
|
| 164 |
+
# Configurer (optionnel pour les presets cloud)
|
| 165 |
cp .env.example .env
|
| 166 |
# Ajouter vos clés API dans .env
|
| 167 |
|
| 168 |
# Lancer avec un preset
|
| 169 |
+
python run_benchmark.py --preset model --verbose
|
| 170 |
|
| 171 |
# Lancer sur un modèle local
|
| 172 |
python run_benchmark.py --endpoint http://localhost:11434/v1/chat/completions \
|
| 173 |
--model yawo-v10 --verbose
|
| 174 |
|
| 175 |
# Évaluer une seule catégorie
|
| 176 |
+
python run_benchmark.py --preset model --category cultural_knowledge -v
|
| 177 |
|
| 178 |
# Comparer deux modèles
|
| 179 |
+
python run_benchmark.py --compare results/model.json results/yawo.json
|
| 180 |
|
| 181 |
# Voir le classement
|
| 182 |
python run_benchmark.py --leaderboard
|
|
|
|
| 198 |
| 10 | Robustesse | 10 | 5% | Cohérence face aux entrées adverses |
|
| 199 |
| | **Total** | **107** | **100%** | |
|
| 200 |
|
| 201 |
+
### Scoring ÈwéScore
|
| 202 |
|
| 203 |
L'**ÈwéScore** est un nombre unique (0-100) représentant la capacité globale en Ewe :
|
| 204 |
|
|
|
|
| 214 |
|
| 215 |
| # | Model | ÈwéScore | Tests Passed | Date |
|
| 216 |
|---|-------|----------|--------------|------|
|
| 217 |
+
| 🥇 | *En attente de soumissions* | - | - | - |
|
| 218 |
|
| 219 |
> **Soumettre vos résultats** : Exécutez le benchmark, puis ouvrez une PR avec votre fichier de résultats dans `results/`.
|
| 220 |
|
|
|
|
| 258 |
We welcome contributions! See [docs/CONTRIBUTING.md](docs/CONTRIBUTING.md) for details.
|
| 259 |
|
| 260 |
Ways to contribute:
|
| 261 |
+
- **Add tests** More tests improve coverage
|
| 262 |
+
- **Validate translations** Native speaker review
|
| 263 |
+
- **Submit results** Run on your model and share
|
| 264 |
+
- **Report issues** Found a bad test? Let us know
|
| 265 |
|
| 266 |
---
|
| 267 |
|
| 268 |
## License
|
| 269 |
|
| 270 |
+
**CC BY-NC 4.0** Creative Commons Attribution-NonCommercial 4.0 International
|
| 271 |
|
| 272 |
- ✅ Free to use for research, education, and evaluation
|
| 273 |
- ✅ Free to modify and redistribute (with attribution)
|
|
|
|
| 291 |
|
| 292 |
<div align="center">
|
| 293 |
|
| 294 |
+
**Created by [Joel Elisée ADZONYA](https://joel.adzonya.strivenew.com) [Strive AI](https://github.com/joel710)**
|
| 295 |
|
| 296 |
+
*L'IA au service des langues africaines*
|
| 297 |
|
| 298 |
</div>
|